انجمن کامپیوتر ایران

Noise robustness in speaker recognition has attracted a great deal of interest. This paper describes a novel technique for noise robust Speaker Identification. In this technique, initially, the speech autocorrelation sequence is computed and then, the effect of noise is suppressed using a high pass filter in autocorrelation domain. Finally, the speech feature set is found using the spectral peaks of this filtered autocorrelation sequence. These Features are robust for speech recognition task. In this paper we applied them to speaker identification task and found that this features are more robust than MFCC features. For example in a test of Farsdat speech database, after 10dB corruption of speech signal using babble noise, it was observed that these features decrease the error rate for more than 22%.